论文部分内容阅读
提出了2种解决汉语语音识别中声调问题的方法:利用区分性方法对基于隐马尔可夫模型(HMM)的声调模型进行训练;提出将区分性训练的声调模型加入大词汇量连续语音识别系统的最优方法,该方法根据最小音子错误的训练准则以及利用扩展Baum-Welch算法区分性训练与模型相关的概率权重,对声学模型以及声调模型概率进行加权.实验结果表明区分性训练的声调模型能够显著地提高连续语音声调识别率以及大词汇量语音识别系统的识别率,同时区分性的模型权重训练能够在区分性声调模型加入连续语音识别系统之后进一步提高系统的识别性能.
Two kinds of methods to solve the tone problem in Chinese speech recognition are proposed: Discriminant method is used to train the HMM-based tone model; Discriminative training tone model is proposed to be added into the Vocabulary Large-Scale Continuous Speech Recognition System The method optimizes the acoustic model and the probability of the tone model according to the training criterion of the smallest pitch error and the probability weight of model-related training using the extended Baum-Welch algorithm. The experimental results show that the tone of the discriminative training The model can significantly improve the rate of continuous voice recognition and the recognition rate of large vocabulary speech recognition systems. Meanwhile, the discriminative model weight training can further improve the recognition performance of the system after the discriminative model is added to the continuous speech recognition system.